bambi

New features

• Add configuration facilities to Bambi (#745)
• Interpet submodule now outputs informative messages when computing default values (#745)
• Bambi supports weighted responses (#761)
• Bambi supports constrained responses (#764)
• Implement compute_log_likelihood() method to compute the log likelihood on a model (#769)
• Add a class InferenceMethods that allows users to access the available inference methods and kwargs (#795)

Maintenance and fixes

• Fix bug in predictions with models using HSGP (#780)
• Fix get_model_covariates() utility function (#801)
• Use pm.compute_deterministics() to compute deterministics when bayeux based samplers are used (#803)
• Wrap all the parameters of the response distribution (the likelihood) with a pm.Deterministic (#804)
• Keep bayeux-ml as the single direct JAX-related dependency (#804)
• The response component only holds response information about the response, not about predictors of the parent parameter (#804)
• Resolve import error associated with bayeux (#822)

Documentation

• Our Code of Conduct now includes how to send a report (#783)
• Add polynomial regression example (#809)
• Add Contact form to our webpage (#816)

Deprecation

• f"{response_name}_obs" has been replaced by "__obs__" as the dimension name for the observation index (#804)
• f"{response_name}_{parameter_name}" is no longer the name for the name of parameters of the likelihood. Now Bambi uses "{parameter_name}" (#804)
• kind in Model.predict() now use "response_params" and "response" instead of "mean" and "pps" (#804)
• include_mean has been replaced by include_response_params in Model.fit() (#804)

This is the first version of Bambi that is released with a Governance structure. Added in #709.
The highlights are the shiny interpret subpackage and the implementation of support for censored models.

New features

• Bambi now supports censored responses (#697)
• Implement "exponential" and "weibull" families (#697)
• Add "kidney" dataset (#697)
• Add interpret submodule (#684, #695, #699, #701, #732, #736)
  • Implements comparisons, predictions, slopes, plot_comparisons, plot_predictions, and plot_slopes
• Support censored families

Maintenance and fixes

• Replace univariate_ordered with ordered (#724)
• Add missing docstring for center_predictors (#726)
• Fix bugs in plot_comparison (#731)

Documentation

• Add docstrings to utility functions (#696)
• Migrate documentation to Quarto (#712)
• Add case study for MRP (#716)
• Add example about ordinal regression (#719)
• Add example about zero inflated models (#725)
• Add example about predictions for new groups (#734)

Deprecation

• Drop official suport for Python 3.8 (#720)
• Change plots submodule name to interpret (#705)

0.12.0

New features

• Implement new families "ordinal" and "sratio" for modeling of ordinal responses (#678)
• Allow families to implement a custom create_extra_pps_coord() (#688)
• Allow predictions on new groups (#693)

Maintenance and fixes

• Robustify how Bambi handles dims (#682)
• Fix links in FAQ (#686)
• Update additional dependencies install command (#689)
• Update predict pps docstring (#690)
• Add warning for aliases athat aren't used (#691)

Documentation

• Add families to the Getting Started guide (#683)

0.11.0

New features

• Add support for Gaussian Processes via the HSGP approximation (#632)
• Add new families: "zero_inflated_poisson", "zero_inflated_binomial", and "zero_inflated_negativebinomial" (#654)
• Add new families: "beta_binomial" and "dirichlet_multinomial" (#659)
• Allow plot_cap() to show predictions at the observation level (#668)
• Add new families: "hurdle_gamma", "hurdle_lognormal", "hurdle_negativebinomial", and "hurdle_poisson" (#676)

Maintenance and fixes

• Modify how HSGP is built in PyMC when there are groups (#661)
• Modify how Bambi is imported in the tests (#662)
• Prevent underscores from being removed in dim names (#664)
• Bump sphinx dependency to a version greater than 7 (#672)

Documentation

• Document how to use custom priors (#656)
• Fix name of arviz traceplot function in the docs (#666)
• Add example that shows how plot_cap() works (#670)

New features

• Refactored the codebase to support distributional models (#607)
• Added a default method to handle posterior predictive sampling for custom families (#625)
• plot_cap() gains a new argument target that allows to plot different parameters of the response distribution (#627)

Maintenance and fixes

• Moved the tests directory to the root of the repository (#607)
• Don't pass dims to the response of the likelihood distribution anymore (#629)
• Remove requirements.txt and replace with pyproject.toml config file to distribute the package (#631)

Documentation

• Update examples to work with the new internals (#607)
• Fixed figure in the Sleepstudy example (#607)
• Add example using distributional models (#641)

Deprecation

• Removed versioned documentation webpage (#616)
• Removed correlated priors for group-specific terms (#607)
• Dictionary with tuple keys are not allowed for priors anymore (#607)

Maintenance and fixes

• Update to PyMC >= 5, which means we use PyTensor instead of Aesara now (#613, #614)

New features

• Implement censored() (#581)
• Add Formula class (#585)
• Add common numpy transforms to extra_namespace (#589)
• Add AsymmetricLaplace family for Quantile Regression (#591)
• Add 'transforms' argument to plot_cap() (#594)
• Add panel covariates to plot_cap() and make it more flexible (#596)

Maintenance and fixes

• Reimplemented predictions to make better usage of xarray data structures (#573)
• Keep 0 dimensional parameters as 0 dimensional instead of 1 dimensional (#575)
• Refactor terms for modularity and extensibility (#582)
• Remove seed argument from model.initial_point() (#592)
• Add build check function on prior predictive and plot prior (#605)

Documentation

• Add quantile regression example (#608)

Deprecation

• Remove automatic_priors argument from Model (#603)
• Remove string option for data input in Model (#604)

Bambi 0.9.1

New features

• Add support for jax sampling via numpyro and blackjax samplers (#526)
• Add Laplace family (#524)
• Improve Laplace computation and integration (#555 and #563)

Maintenance and fixes

• Ensure order variable is preserved when ploting priors (#529)
• Treat offset accordingly (#534)
• Refactor tests to share data generation code (#531)

Documentation

• Update documentation following good inferencedata practices (#537)
• Add logos to repo and docs (#542)

Deprecation

• Deprecate method argument in favor of inference_method (#554)

New features

• Bambi now uses PyMC 4.0 as it's backend. Most if not all your previous model should run the same, without the need of any change.
• Add Plot Conditional Adjusted Predictions plot_cap (#517)

Maintenance and fixes

• Group specific terms now work with numeric of multiple columns (#516)

Bambi 0.8.0

New features

• Add VonMises ("vonmises") built-in family (#453)
• Model.predict() gains a new argument include_group_specific to determine if group-specific effects are considered when making predictions (#470)
• Add Multinomial ("multinomial") built-in family (#490)

Maintenance and fixes

• Add posterior predictive sampling method to "categorical" family (#458)
• Require Python >= 3.7.2 to fix NoReturn type bug in Python (#463)
• Fixed the wrong builtin link given by link="inverse" was wrong. It returned the same result as link="cloglog" (#472)
• Replaced plain dictionaries with namedtuples when same dictionary structure was repeated many times (#472)
• The function check_full_rank() in utils.py now checks the array is 2 dimensional (#472)
• Removed _extract_family_prior() from bambi/families as it was unnecesary (#472)
• Removed bambi/families/utils.py as it was unnecessary (#472)
• Removed external links and unused datasets (#483)
• Replaced "_coord_group_factor" with "__factor_dim" and "_coord_group_expr" with "__expr_dim" in dimension/coord names (#499)
• Fixed a bug related to modifying the types of the columns in the original data frame (#502)

Documentation

• Add circular regression example (#465)
• Add Categorical regression example (#457)
• Add Beta regression example (#442)
• Add Radon Example (#440)
• Fix typos and clear up writing in some docs (#462)
• Documented the module bambi/defaults (#472)
• Improved documentation and made it more consistent (#472)
• Cleaned Strack RRR example (#479)

Deprecation

• Removed old default priors (#474)
• Removed draws parameter from Model.predict() method (#504)

This is a patch release where we fix a bug related to the shape of 2 level categorical group-specific effects (#441)

This release includes a mix of new features, fixes, and new examples on our webpage.

New features

• Add "categorical" built-in family (#426)
• Add include_mean argument to the method Model.fit() (#434)
• Add .set_alias() method to Model (#435)

Maintenance and fixes

• Codebase for the PyMC backend has been refactored (#408)
• Fix examples that averaged posterior values across chains (#429)
• Fix issue #427 with automatic priors for the intercept term (#430)

Documentation

• Add StudentT regression example, thanks to @tjburch (#414)
• Add B-Spline regression example with cherry blossoms dataset (#416)
• Add hirarchical linear regression example with sleepstudy dataset (#424)

Use formulae 0.2.0

Minor fixes to code and docs

Mainly changes to the docs and minor fixes.

Many changes are included in this release. Some of the most important changes are

• New model families (StudentT, Binomial, Beta).
• In-sample and out-of-sample predictions.
• Improved sampling performance due to predictor centering when the model contains an intercept.
• New default priors (similar to rstanarm default priors).
• It's possible to use potentials.
• There's a new function to load datasets used throughout examples

The main changes in this release can be summarized as follows

• Modified the API. Now all information relative to the model is passed in Model instantiation instead of in Model.fit().
• Fixed Gamma, Wald, and Negative Binomial families.
• Changed theme of the webpage and now the documentation is built automatically.

The aim of this release is to update to formulae 0.0.9, which contains several bug fixes. There are also other minor fixes and improvements that can be found in the changelog.

The main change in this release is the use of formulae, instead of patsy, to parse model formulas.